Given Bayes neural network method, the direct marketing customer repurchase behavior shall be built to analyze how different customer groups have different expectations on ATM services. Also, application of GMKL algorithm in customer recognition of network finance is proposed, training with the use of 20, 000 customer records extracted from RESSET/DB financial research database. First, transform the numerical characteristics of sample for collected customer record and normalize the data. Next, try to obtain multi-kernel function with different array modes of kernel function and compare the results, and the results show that accuracy of MKL in customer recognition of network finance can be up to 86. 4%, which is obvious higher than that of support vector machine